[FieldTrip] using cfg.runica.pca to reduce number of ICs

Tue Apr 8 17:45:19 CEST 2014

Hi Fred,

This is a difficult question and I don't know the answer. Searching a bit more, I've found that some experts strongly do not recommend PCA before ICA with very good reasons:

http://sccn.ucsd.edu/pipermail/eeglablist/2010/003339.html
http://sccn.ucsd.edu/pipermail/eeglablist/2013/006101.html

Another possibility (the one I'm using in my data) is to correlate the vertical EOG time courses (or the horizontal EOG) with all ICs. The rationale is that you want to find the independent component whose time courses are more similar to the EOG time courses. Then you can reduce the potential IC candidates that contain the prototypical artifacts using prior knowledge.

You'll see that 2/3 of them show high correlations (in comparison with other ICs) with topographies that resemble blinks (or eye movements respectively). For ECG you can try the same strategy or use a very similar approach using coherence:
http://fieldtrip.fcdonders.nl/example/use_independent_component_analysis_ica_to_remove_ecg_artifacts

However, you should had recorded the ECG, vertical and horizontal EOGs, butI don't know if this is the case. If not, I'd recommend you to record them in future experiments (just by my own experience).

Best,

Diego

----- Original Message -----
> From: "Frédéric Roux" <f.roux at bcbl.eu>
> To: "FieldTrip discussion list" <fieldtrip at science.ru.nl>
> Sent: Tuesday, 8 April, 2014 4:06:36 PM
> Subject: Re: [FieldTrip] using cfg.runica.pca to reduce number of ICs
> Hi Diego,
> 
> thanks for the quick reply.
> 
> When I compute the rank of the concatenated trials I get rank(concat1)
> = 202, which is the number of
> channels that are in the data (planar gradiometers only). So in fact
> that number corresponds to the default
> output returned by ft_componentanalysis.
> 
> Alos, I usually run ICA without PCA component reduction and can
> identify EOG and ECG quite easily by
> eye-balling. But, I'd like to figure out what the advantages of PCA
> reduction are.
> 
> For instance, will reducing the number of ICs through PCA help to
> isolate better EOG and ECG components
> or will the decomposition be the same the only difference being that
> the algorithm will run faster?
> 
> Best,
> 
> Fred
> 
> Frédéric Roux
> 
> ----- Original Message -----
> From: "Mauricio Antelis" <mauricio.antelis at gmail.com>
> To: "Diego Lozano" <d.lozanosoldevilla at fcdonders.ru.nl>, "FieldTrip
> discussion list" <fieldtrip at science.ru.nl>
> Sent: Tuesday, April 8, 2014 3:36:42 PM
> Subject: Re: [FieldTrip] using cfg.runica.pca to reduce number of ICs
> 
> 
> 
> Roberto,
> 
> 
> Aqui algo sobre el cuentionamiento del numero de componentes para
> realizar ICA, sin embargo en nuestro caso, creo que no sera un
> parametro sensible ya que tenemos un numero bajo (21) de mediciones
> 
> 
> Saludos
> 
> 
> Mauricio
> 
> 
> 
> 
> 
> On Tue, Apr 8, 2014 at 8:29 AM, Lozano Soldevilla, D. (Diego) <
> d.lozanosoldevilla at fcdonders.ru.nl > wrote:
> 
> 
> Hi Fred,
> 
> I don't know the magical number but I see the following options:
> 
> a) Before ICA, concatenate all single trials and ask for the rank of
> your data. Use it as your cfg.runica.pca input.
> 
> b) Some users notice that when you sort IC by variance, beyond
> component 30, the IC topo/activation does not look like
> physiologically meaningfull. It matches with the PCA reduction before
> ICA with other papers:
> 
> http://www.ncbi.nlm.nih.gov/pubmed/15219593
> http://www.ncbi.nlm.nih.gov/pubmed/19699307
> 
> Might be somebody in the forum have tried (by simulations or in real
> data) on the effects of PCA component reduction on ICA.
> 
> I hope it helps,
> 
> Diego
> 
> ----- Original Message -----
> > From: "Frédéric Roux" < f.roux at bcbl.eu >
> > To: "FieldTrip discussion list" < fieldtrip at science.ru.nl >
> > Sent: Tuesday, 8 April, 2014 2:46:22 PM
> > Subject: [FieldTrip] using cfg.runica.pca to reduce number of ICs
> > Dear all,
> >
> > I have a general question relating to the usage of cfg.runica.pca
> > during the call to ft_componentanalysis.
> >
> > Is it correct to assume that cfg.runica.pca = length(meg_data.label)
> > will
> > force the algorithm to return n = length(meg_data.label) ICs, and
> > that
> > as a
> > result artifacts can be "spread" across several ICs?
> >
> > If that's true, then I imagine that cfg.runica.pac = n/4 will return
> > less components
> > and reduce the "spread" of artifacts over several components.
> >
> > My question is how to choose the number of principal components to
> > which the data
> > is reduced before ICA?
> >
> > Best,
> > Fred
> > ---------------------------------------------------------------------------
> >
> >
> > _______________________________________________
> > fieldtrip mailing list
> > fieldtrip at donders.ru.nl
> > http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
> 
> --
> PhD Student
> Neuronal Oscillations Group
> Donders Institute for Brain, Cognition and Behaviour
> Centre for Cognitive Neuroimaging
> Radboud University Nijmegen
> NL-6525 EN Nijmegen
> The Netherlands
> http://www.ru.nl/people/donders/lozano-soldevilla-d/
> 
> _______________________________________________
> fieldtrip mailing list
> fieldtrip at donders.ru.nl
> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
> 
> 
> 
> --
> 
> Javier M. Antelis
> 
> _______________________________________________
> fieldtrip mailing list
> fieldtrip at donders.ru.nl
> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
> 
> _______________________________________________
> fieldtrip mailing list
> fieldtrip at donders.ru.nl
> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip

-- 
PhD Student
Neuronal Oscillations Group
Donders Institute for Brain, Cognition and Behaviour 
Centre for Cognitive Neuroimaging
Radboud University Nijmegen 
NL-6525 EN Nijmegen
The Netherlands
http://www.ru.nl/people/donders/lozano-soldevilla-d/